Approach 1

Overall

For each lion's class, extract samples, build a dataset and train a convnet. Then search and count each lion's class findings on target images using a sliding window.

Steps

1. Explore and validate dataset (dotted samples vs csv counting)
2. Export lion samples to dataset
   • For each image on training folder:
     1. Find dots by class (through dot color)
     2. Filter out animals near than 50px from each other (try to avoid noisy images due to overlapping)
     3. Export 124x124 images centered on dots to a dataset, labeled with lion's class (1-5) in one hot encoding
        • Ideas: ignore lions that are too near each other to minimize data noise?
3. Adjust lion classes distribution on dataset
4. Train dataset to classify lion classes
5. Test: Find and count lion classes on train images and compare to csv class count
   1. Generate object candidates (lets avoid a full sliding window for search)
   2. For each object candidate, predict for a lion class. If good, count class match. Avoid recount overlappings.
   3. Show error count by class and generate RMSE
6. Generate submission file